Forecast demand across 30,490 product-store series

To set ordering, inventory, staffing, and transport plans, a retailer must forecast demand across products, stores, categories, and aggregate planning levels. Error at any level can create lost sales, waste, excess working capital, or avoidable logistics cost.

The case covers 30,490 product-store series and a 28-day horizon. It therefore tests accuracy and consistency across the hierarchy rather than one aggregate series.

To compare methods, the case uses weighted root mean squared scaled error (WRMSSE), which assigns more weight to products with higher sales value. The benchmark compares eomer with a traditional pipeline that uses hand-built features.

The runtime test covers the same nationwide series set. The value model then applies explicit assumptions for stockout loss, waste, recovered value, and data-science time.

The case reports a WRMSSE of 0.738 for eomer and 0.765 for the traditional benchmark. The eomer forecast completes in about three minutes without a hand-built feature set, compared with about nine minutes for the benchmark pipeline.

The USD 5.6 million annual value estimate is illustrative and depends on the stated assumptions. A deployment should replace these assumptions with observed stockout, markdown, waste, labour, and revenue data.